GCC Middle and Back End API Reference
gcse.c File Reference
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "diagnostic-core.h"
#include "toplev.h"
#include "hard-reg-set.h"
#include "rtl.h"
#include "tree.h"
#include "tm_p.h"
#include "regs.h"
#include "ira.h"
#include "flags.h"
#include "insn-config.h"
#include "recog.h"
#include "basic-block.h"
#include "function.h"
#include "expr.h"
#include "except.h"
#include "ggc.h"
#include "params.h"
#include "cselib.h"
#include "intl.h"
#include "obstack.h"
#include "tree-pass.h"
#include "hash-table.h"
#include "df.h"
#include "dbgcnt.h"
#include "target.h"
#include "gcse.h"
#include "gt-gcse.h"
Include dependency graph for gcse.c:

Data Structures

struct  expr
struct  occr
struct  hash_table_d
struct  ls_expr
struct  pre_ldst_expr_hasher
struct  modify_pair_s
struct  bb_data
struct  reg_avail_info
struct  mem_conflict_info

Macros

#define BB_DATA(bb)   ((struct bb_data *) (bb)->aux)
#define GNEW(T)   ((T *) gmalloc (sizeof (T)))
#define GCNEW(T)   ((T *) gcalloc (1, sizeof (T)))
#define GNEWVEC(T, N)   ((T *) gmalloc (sizeof (T) * (N)))
#define GCNEWVEC(T, N)   ((T *) gcalloc ((N), sizeof (T)))
#define GNEWVAR(T, S)   ((T *) gmalloc ((S)))
#define GCNEWVAR(T, S)   ((T *) gcalloc (1, (S)))
#define GOBNEW(T)   ((T *) gcse_alloc (sizeof (T)))
#define GOBNEWVAR(T, S)   ((T *) gcse_alloc ((S)))
#define can_copy   (this_target_gcse->x_can_copy)
#define can_copy_init_p   (this_target_gcse->x_can_copy_init_p)

Typedefs

typedef struct occroccr_t
typedef struct modify_pair_s modify_pair

Functions

static int expr_equiv_p (const_rtx, const_rtx)
static void compute_can_copy (void)
static void * gmalloc (static void *gcalloc(size_t)
bool can_copy_p ()
static void * gmalloc ()
static void * gcalloc ()
static void * gcse_alloc ()
static void alloc_gcse_mem ()
static void free_gcse_mem ()
static void compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc, struct hash_table_d *table)
static int want_to_gcse_p ()
bool can_assign_to_reg_without_clobbers_p ()
static int oprs_unchanged_p ()
static void mems_conflict_for_gcse_p (rtx dest, const_rtx setter, void *data)
static int load_killed_in_block_p (const_basic_block bb, int uid_limit, const_rtx x, int avail_p)
static int oprs_anticipatable_p ()
static int oprs_available_p ()
static unsigned int hash_expr (const_rtx x, enum machine_mode mode, int *do_not_record_p, int hash_table_size)
static int expr_equiv_p ()
static void insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p, int avail_p, int max_distance, struct hash_table_d *table)
static void hash_scan_set ()
static void hash_scan_clobber (rtx x, rtx insn, struct hash_table_d *table)
static void hash_scan_call (rtx x, rtx insn, struct hash_table_d *table)
static void hash_scan_insn ()
static void dump_hash_table ()
static void record_last_reg_set_info ()
static void canon_list_insert (rtx dest, const_rtx x, void *v_insn)
static void record_last_mem_set_info ()
static void record_last_set_info ()
static void compute_hash_table_work ()
static void alloc_hash_table ()
static void free_hash_table ()
static void compute_hash_table ()
static void clear_modify_mem_tables ()
static void free_modify_mem_tables ()
static void compute_transp ()
static void alloc_pre_mem ()
static void free_pre_mem ()
static void prune_expressions ()
static void prune_insertions_deletions ()
static struct edge_listcompute_pre_data ()
static int pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited)
static int pre_expr_reaches_here_p ()
static rtx process_insert_insn ()
static void insert_insn_end_basic_block ()
static int pre_edge_insert ()
static void pre_insert_copy_insn ()
static void pre_insert_copies ()
static rtx gcse_emit_move_after ()
static int pre_delete ()
static int pre_gcse ()
static int one_pre_gcse_pass ()
static void add_label_notes ()
static void alloc_code_hoist_mem ()
static void free_code_hoist_mem ()
static void compute_code_hoist_vbeinout ()
static void compute_code_hoist_data ()
static int update_bb_reg_pressure ()
static int should_hoist_expr_to_dom (basic_block expr_bb, struct expr *expr, basic_block bb, sbitmap visited, int distance, int *bb_size, enum reg_class pressure_class, int *nregs, bitmap hoisted_bbs, rtx from)
static struct occrfind_occr_in_bb ()
static int hoist_code ()
static enum reg_class get_regno_pressure_class ()
static enum reg_class get_pressure_class_and_nregs ()
static void change_pressure ()
static void calculate_bb_reg_pressure ()
static int one_code_hoisting_pass ()
static struct ls_exprldst_entry ()
static void free_ldst_entry ()
static void free_ld_motion_mems ()
static void print_ldst_list ()
static struct ls_exprfind_rtx_in_ldst ()
static int simple_mem ()
static void invalidate_any_buried_refs ()
static void compute_ld_motion_mems ()
static void trim_ld_motion_mems ()
static void update_ld_motion_stores ()
static bool is_too_expensive ()
static bool gate_rtl_pre ()
static unsigned int execute_rtl_pre ()
static bool gate_rtl_hoist ()
static unsigned int execute_rtl_hoist ()
rtl_opt_passmake_pass_rtl_pre ()
rtl_opt_passmake_pass_rtl_hoist ()

Variables

struct target_gcse default_target_gcse
int flag_rerun_cse_after_global_opts
static struct obstack gcse_obstack
static struct hash_table_d expr_hash_table
static struct ls_exprpre_ldst_mems = NULL
static hash_table
< pre_ldst_expr_hasher
pre_ldst_table
static regset reg_set_bitmap
static vec< rtx > * modify_mem_list
static bitmap modify_mem_list_set
static vec< modify_pair > * canon_modify_mem_list
static bitmap blocks_with_calls
static int bytes_used
static int gcse_subst_count
static int gcse_create_count
static bool doing_code_hoisting_p = false
static sbitmapae_kill
static basic_block curr_bb
static int curr_reg_pressure [N_REG_CLASSES]
static struct reg_avail_inforeg_avail_info
static basic_block current_bb
static rtx test_insn
static sbitmaptransp
static sbitmapcomp
static sbitmapantloc
static sbitmappre_optimal
static sbitmappre_redundant
static sbitmappre_insert_map
static sbitmappre_delete_map
static sbitmaphoist_vbein
static sbitmaphoist_vbeout

Macro Definition Documentation

#define BB_DATA (   bb)    ((struct bb_data *) (bb)->aux)
#define can_copy   (this_target_gcse->x_can_copy)

Referenced by gmalloc().

#define can_copy_init_p   (this_target_gcse->x_can_copy_init_p)
#define GCNEW (   T)    ((T *) gcalloc (1, sizeof (T)))
#define GCNEWVAR (   T,
  S 
)    ((T *) gcalloc (1, (S)))
#define GCNEWVEC (   T,
  N 
)    ((T *) gcalloc ((N), sizeof (T)))

Referenced by can_copy_p().

#define GNEW (   T)    ((T *) gmalloc (sizeof (T)))
#define GNEWVAR (   T,
  S 
)    ((T *) gmalloc ((S)))

Referenced by canon_list_insert().

#define GNEWVEC (   T,
  N 
)    ((T *) gmalloc (sizeof (T) * (N)))
#define GOBNEW (   T)    ((T *) gcse_alloc (sizeof (T)))
#define GOBNEWVAR (   T,
  S 
)    ((T *) gcse_alloc ((S)))

Typedef Documentation

typedef struct modify_pair_s modify_pair
typedef struct occr* occr_t

Function Documentation

static void add_label_notes ( )
static

If X contains any LABEL_REF's, add REG_LABEL_OPERAND notes for them to INSN. If such notes are added to an insn which references a CODE_LABEL, the LABEL_NUSES count is incremented. We have to add that note, because the following loop optimization pass requires them. ??? If there was a jump optimization pass after gcse and before loop, then we would not need to do this here, because jump would add the necessary REG_LABEL_OPERAND and REG_LABEL_TARGET notes.

This code used to ignore labels that referred to dispatch tables to avoid flow generating (slightly) worse code.

We no longer ignore such label references (see LABEL_REF handling in mark_jump_label for additional information).

     There's no reason for current users to emit jump-insns with
     such a LABEL_REF, so we don't have to handle REG_LABEL_TARGET
     notes.   

References bb_data::backup, BB_DATA, bitmap_bit_p, bitmap_copy(), basic_block_def::index, bb_data::live_in, bb_data::max_reg_pressure, bb_data::old_pressure, and update_bb_reg_pressure().

Referenced by gcse_emit_move_after().

static void alloc_code_hoist_mem ( )
static

??? We could compute post dominators and run this algorithm in reverse to perform tail merging, doing so would probably be more effective than the tail merging code in jump.c.

It's unclear if tail merging could be run in parallel with code hoisting. It would be nice. Allocate vars used for code hoisting analysis.

static void alloc_gcse_mem ( )
static

Allocate memory for the reg/memory set tracking tables. This is called at the start of each pass.

Allocate vars to track sets of regs.

 Allocate array to keep a list of insns which modify memory in each
 basic block.  The two typedefs are needed to work around the
 pre-processor limitation with template types in macro arguments.   

References bitmap_vector_clear(), bitmap_vector_ones(), expr::expr, last_basic_block, and hash_table_d::size.

Referenced by pre_insert_copies().

static void alloc_hash_table ( )
static

Allocate space for the set/expr hash TABLE. It is used to determine the number of buckets to use.

Attempt to maintain efficient use of hash table. Making it an odd number is simplest for now. ??? Later take some measurements.

Referenced by pre_insert_copies().

static void alloc_pre_mem ( )
static

Allocate vars used for PRE analysis.

pre_insert and pre_delete are allocated later.

References EXECUTE_IF_SET_IN_BITMAP.

Referenced by pre_insert_copies().

static void calculate_bb_reg_pressure ( )
static

Calculate register pressure for each basic block by walking insns from last to first.

References cfun, FLOAT_MODE_P, GET_MODE, may_trap_p(), MEM_VOLATILE_P, reg_mentioned_p(), side_effects_p(), and stack_pointer_rtx.

bool can_assign_to_reg_without_clobbers_p ( )

Return true if we can assign X to a pseudo register such that the resulting insn does not result in clobbering a hard register as a side-effect.

Additionally, if the target requires it, check that the resulting insn can be copied. If it cannot, this means that X is special and probably has hidden side-effects we don't want to mess with.

This function is typically used by code motion passes, to verify that it is safe to insert an insn without worrying about clobbering maybe live hard regs.

 If this is a valid operand, we are OK.  If it's VOIDmode, we aren't.   
 Otherwise, check if we can make a valid insn from it.  First initialize
 our test insn if we haven't already.   
 Now make an insn like the one we would make when GCSE'ing and see if
 valid.   

Referenced by one_code_hoisting_pass().

bool can_copy_p ( )

Returns whether the mode supports reg/reg copy operations.

References GCNEWVEC.

Referenced by reset_opr_set_tables().

static void canon_list_insert ( rtx  dest,
const_rtx  x,
void *  v_insn 
)
static

Record all of the canonicalized MEMs of record_last_mem_set_info's insn. Note we store a pair of elements in the list, so they have to be taken off pairwise.

If DEST is not a MEM, then it will not conflict with a load. Note that function calls are assumed to clobber memory, but are handled elsewhere.

References get_max_insn_count(), and GNEWVAR.

static void change_pressure ( )
static

Increase (if INCR_P) or decrease current register pressure for register REGNO.

static void clear_modify_mem_tables ( )
static

Expression tracking support. Clear canon_modify_mem_list and modify_mem_list tables.

Referenced by record_last_set_info().

static void compute_can_copy ( void  )
static
static void compute_code_hoist_data ( )
static

Top level routine to do the dataflow analysis needed by code hoisting.

static void compute_code_hoist_vbeinout ( )
static

Compute the very busy expressions at entry/exit from each block.

An expression is very busy if all paths from a given point compute the expression.

We scan the blocks in the reverse order to speed up the convergence.

             Include expressions in VBEout that are calculated
             in BB and available at its end.   

References bitmap_set_bit, EXECUTE_IF_SET_IN_BITMAP, basic_block_def::index, and sbitmap_free().

static void compute_hash_table ( void  )
static

Compute the expression hash table TABLE.

Initialize count of number of entries in hash table.

Referenced by pre_insert_copies().

static void compute_hash_table_work ( )
static

Top level function to create an expression hash table.

Expression entries are placed in the hash table if

  • they are of the form (set (pseudo-reg) src),
  • src is something we want to perform GCSE on,
  • none of the operands are subsequently modified in the block

Currently src must be a pseudo-reg or a const_int.

TABLE is the table computed.

 re-Cache any INSN_LIST nodes we have allocated.   
 Some working arrays used to track first and last set in each block.   
     First pass over the instructions records information used to
     determine when registers and memory are first and last set.   
     The next pass builds the hash table.   
static void compute_ld_motion_mems ( )
static

Find all the 'simple' MEMs which are used in LOADs and STORES. Simple being defined as MEM loads and stores to symbols, with no side effects and no registers in the expression. For a MEM destination, we also check that the insn is still valid if we replace the destination with a REG, as is done in update_ld_motion_stores. If there are any uses/defs which don't match this criteria, they are invalidated and trimmed out later.

                 Check for a simple LOAD...   
                     Make sure there isn't a buried load somewhere.   
                 Check for stores. Don't worry about aliased ones, they
                 will block any movement we might do later. We only care
                 about this exact pattern since those are the only
                 circumstance that we will ignore the aliasing info.   
                         Check for REG manually since want_to_gcse_p
                         returns 0 for all REGs.   

References execute(), execute_rtl_hoist(), and gate_rtl_hoist().

Referenced by pre_insert_copies().

static void compute_local_properties ( sbitmap transp,
sbitmap comp,
sbitmap antloc,
struct hash_table_d table 
)
static

Compute the local properties of each recorded expression.

Local properties are those that are defined by the block, irrespective of other blocks.

An expression is transparent in a block if its operands are not modified in the block.

An expression is computed (locally available) in a block if it is computed at least once and expression would contain the same value if the computation was moved to the end of the block.

An expression is locally anticipatable in a block if it is computed at least once and expression would contain the same value if the computation was moved to the beginning of the block.

We call this routine for pre and code hoisting. They all compute basically the same information and thus can easily share this code.

TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local properties. If NULL, then it is not necessary to compute or record that particular property.

TABLE controls which hash table to look at.

 Initialize any bitmaps that were passed in.   
         The expression is transparent in this block if it is not killed.
         We start by assuming all are transparent [none are killed], and
         then reset the bits for those that are.   
         The occurrences recorded in antic_occr are exactly those that
         we want to set to nonzero in ANTLOC.   
               While we're scanning the table, this is a good place to
               initialize this.   
         The occurrences recorded in avail_occr are exactly those that
         we want to set to nonzero in COMP.   
               While we're scanning the table, this is a good place to
               initialize this.   
         While we're scanning the table, this is a good place to
         initialize this.   
static struct edge_list* compute_pre_data ( )
staticread

Top level routine to do the dataflow analysis needed by PRE.

Compute ae_kill for each basic block using:

~(TRANSP | COMP)

Referenced by pre_insert_copies().

static void compute_transp ( )
static

For each block, compute whether X is transparent. X is either an expression or an assignment [though we don't care which, for this context an assignment is treated as an expression]. For each block where an element of X is modified, reset the INDX bit in BMAP.

 repeat is used to turn tail-recursion into iteration since GCC
 can't do it when there's no return value.   
         First handle all the blocks with calls.  We don't need to
         do any list walking for them.   
         Now iterate over the blocks which have memory modifications
         but which do not have any calls.   
         If we are about to do the last recursive call
         needed at this level, change it into iteration.
         This function is called enough to be worth it.   

References NULL, and sbitmap_vector_free().

Referenced by free_gcse_mem().

static void dump_hash_table ( )
static

Dump the hash table TABLE to file FILE under the name NAME.

Flattened out table, so it's printed in proper order.

Referenced by pre_insert_copies().

static unsigned int execute_rtl_hoist ( )
static

Referenced by compute_ld_motion_mems().

static unsigned int execute_rtl_pre ( )
static
static int expr_equiv_p ( const_rtx  ,
const_rtx   
)
static
static int expr_equiv_p ( )
static

Return nonzero if exp1 is equivalent to exp2.

static struct occr* find_occr_in_bb ( )
staticread

Find occurrence in BB.

Find the right occurrence of this expression.

References bb_data::backup, BASIC_BLOCK, BB_DATA, bitmap_copy(), bb_data::live_in, bb_data::max_reg_pressure, and bb_data::old_pressure.

Referenced by should_hoist_expr_to_dom().

static struct ls_expr* find_rtx_in_ldst ( )
staticread

Returns 1 if X is in the list of ldst only expressions.

References cfun, dbg_cnt(), and optimize_function_for_speed_p().

static void free_code_hoist_mem ( )
static

Free vars used for code hoisting analysis.

References bitmap_set_bit, basic_block_def::index, and should_hoist_expr_to_dom().

static void free_gcse_mem ( )
static
static void free_hash_table ( )
static

Free things allocated by alloc_hash_table.

References XEXP.

Referenced by pre_insert_copies().

static void free_ld_motion_mems ( )
static

Free up all memory associated with the ldst list.

Referenced by pre_insert_copies().

static void free_ldst_entry ( )
static

Free up an individual ldst entry.

References dump_file, and print_rtl().

static void free_modify_mem_tables ( )
static

Release memory used by modify_mem_list_set.

References NULL, and sbitmap_vector_alloc().

static void free_pre_mem ( )
static

Free vars used for PRE analysis.

ANTLOC and AE_KILL are freed just after pre_lcm finishes.

Referenced by pre_insert_copies().

static bool gate_rtl_hoist ( )
static

It does not make sense to run code hoisting unless we are optimizing for code size – it rarely makes programs faster, and can make then bigger if we did PRE (when optimizing for space, we don't run PRE).

Referenced by compute_ld_motion_mems().

static bool gate_rtl_pre ( )
static

All the passes implemented in this file. Each pass has its own gate and execute function, and at the end of the file a pass definition for passes.c.

We do not construct an accurate cfg in functions which call setjmp, so none of these passes runs if the function calls setjmp. FIXME: Should just handle setjmp via REG_SETJMP notes.

Referenced by invalidate_any_buried_refs().

static void* gcalloc ( )
static

Cover function to xcalloc to record bytes allocated.

static void* gcse_alloc ( )
static

Cover function to obstack_alloc.

static rtx gcse_emit_move_after ( )
static

Emit move from SRC to DEST noting the equivalence with expression computed in INSN.

This should never fail since we're creating a reg->reg copy we've verified to be valid.

 Note the equivalence for local CSE pass.  Take the note from the old
 set if there was one.  Otherwise record the SET_SRC from the old set
 unless DEST is also an operand of the SET_SRC.   

References add_label_notes(), add_reg_note(), gcc_assert, GET_CODE, GET_RTX_FORMAT, GET_RTX_LENGTH, JUMP_P, LABEL_NUSES, LABEL_P, LABEL_REF_NONLOCAL_P, XEXP, XVECEXP, and XVECLEN.

static enum reg_class get_pressure_class_and_nregs ( )
static

Return pressure class and number of hard registers (through *NREGS) for destination of INSN.

Considered invariant insns have only one set.

static enum reg_class get_regno_pressure_class ( )
static

Return pressure class and number of needed hard registers (through *NREGS) of register REGNO.

Referenced by one_pre_gcse_pass().

static void* gmalloc ( static void *gcalloc size_t)
static

Misc. utilities. Compute which modes support reg/reg copy operations.

References can_copy, emit_insn(), end_sequence(), gen_rtx_REG(), GET_MODE_CLASS, LAST_VIRTUAL_REGISTER, NULL, PATTERN, recog(), and start_sequence().

static void* gmalloc ( )
static

Cover function to xmalloc to record bytes allocated.

static unsigned int hash_expr ( const_rtx  x,
enum machine_mode  mode,
int *  do_not_record_p,
int  hash_table_size 
)
static

Hash expression X.

MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean indicating if a volatile operand is found or if the expression contains something we don't want to insert in the table. HASH_TABLE_SIZE is the current size of the hash table to be probed.

References expr::antic_occr, occr::deleted_p, GOBNEW, occr::insn, and occr::next.

Referenced by load_killed_in_block_p().

static void hash_scan_call ( rtx  x,
rtx  insn,
struct hash_table_d table 
)
static

Currently nothing to do.

static void hash_scan_clobber ( rtx  x,
rtx  insn,
struct hash_table_d table 
)
static

Currently nothing to do.

static void hash_scan_insn ( )
static

Process INSN and add hash table entries as appropriate.

Pick out the sets of INSN and for other forms of instructions record what's been modified.

static void hash_scan_set ( )
static

Scan SET present in INSN and add an entry to the hash TABLE.

     See if a REG_EQUAL note shows this equivalent to a simpler expression.

     This allows us to do a single GCSE pass and still eliminate
     redundant constants, addresses or other expressions that are
     constructed with multiple instructions.

     However, keep the original SRC if INSN is a simple reg-reg move.
     In this case, there will almost always be a REG_EQUAL note on the
     insn that sets SRC.  By recording the REG_EQUAL value here as SRC
     for INSN, we miss copy propagation opportunities and we perform the
     same PRE GCSE operation repeatedly on the same REG_EQUAL value if we
     do more than one PRE GCSE pass.

     Note that this does not impede profitable constant propagations.  We
     "look through" reg-reg sets in lookup_avail_set.   
     Only record sets of pseudo-regs in the hash table.   
         Don't GCSE something if we can't do a reg/reg copy.   
         GCSE commonly inserts instruction after the insn.  We can't
         do that easily for EH edges so disable GCSE on these for now.   
         ??? We can now easily create new EH landing pads at the
         gimple level, for splitting edges; there's no reason we
         can't do the same thing at the rtl level.   
         Is SET_SRC something we want to gcse?   
         Don't CSE a nop.   
         Don't GCSE if it has attached REG_EQUIV note.
         At this point this only function parameters should have
         REG_EQUIV notes and if the argument slot is used somewhere
         explicitly, it means address of parameter has been taken,
         so we should not extend the lifetime of the pseudo.   
         An expression is not anticipatable if its operands are
         modified before this insn or if this is not the only SET in
         this insn.  The latter condition does not have to mean that
         SRC itself is not anticipatable, but we just will not be
         able to handle code motion of insns with multiple sets.   
         An expression is not available if its operands are
         subsequently modified, including this insn.  It's also not
         available if this is a branch, because we can't insert
         a set after the branch.   
 In case of store we want to consider the memory value as available in
 the REG stored in that memory. This makes it possible to remove
 redundant loads from due to stores to the same location.   
       Only record sets of pseudo-regs in the hash table.   
          Don't GCSE something if we can't do a reg/reg copy.   
          GCSE commonly inserts instruction after the insn.  We can't
          do that easily for EH edges so disable GCSE on these for now.   
          Is SET_DEST something we want to gcse?   
          Don't CSE a nop.   
          Don't GCSE if it has attached REG_EQUIV note.
          At this point this only function parameters should have
          REG_EQUIV notes and if the argument slot is used somewhere
          explicitly, it means address of parameter has been taken,
          so we should not extend the lifetime of the pseudo.   
              Stores are never anticipatable.   
              An expression is not available if its operands are
              subsequently modified, including this insn.  It's also not
              available if this is a branch, because we can't insert
              a set after the branch.   
              Record the memory expression (DEST) in the hash table.   
static int hoist_code ( )
static

Actually perform code hoisting.

The code hoisting pass can hoist multiple computations of the same expression along dominated path to a dominating basic block, like from b2/b3 to b1 as depicted below:

   b1      &mdash;&mdash;
   /\         |
  /  \        |
 bx   by   distance
/      \      |

/ \ | b2 b3 ——

Unfortunately code hoisting generally extends the live range of an output pseudo register, which increases register pressure and hurts register allocation. To address this issue, an attribute MAX_DISTANCE is computed and attached to each expression. The attribute is computed from rtx cost of the corresponding expression and it's used to control how long the expression can be hoisted up in flow graph. As the expression is hoisted up in flow graph, GCC decreases its DISTANCE and stops the hoist if DISTANCE reaches 0. Code hoisting can decrease register pressure if live ranges of inputs are shrunk.

Option "-fira-hoist-pressure" implements register pressure directed hoist based on upper method. The rationale is:

  1. Calculate register pressure for each basic block by reusing IRA facility.
  2. When expression is hoisted through one basic block, GCC checks the change of live ranges for inputs/output. The basic block's register pressure will be increased because of extended live range of output. However, register pressure will be decreased if the live ranges of inputs are shrunk.
  3. After knowing how hoisting affects register pressure, GCC prefers to hoist the expression if it can decrease register pressure, by increasing DISTANCE of the corresponding expression.
  4. If hoisting the expression increases register pressure, GCC checks register pressure of the basic block and decrease DISTANCE only if the register pressure is high. In other words, expression will be hoisted through at no cost if the basic block has low register pressure.
  5. Update register pressure information for basic blocks through which expression is hoisted.
 Basic blocks that have occurrences reachable from BB.   
 Basic blocks through which expr is hoisted.   
 Compute a mapping from expression number (`bitmap_index') to
 hash table entry.   
 Calculate sizes of basic blocks and note how far
 each instruction is from the start of its block.  We then use this
 data to restrict distance an expression can travel.   
         Don't count debug instructions to avoid them affecting
         decision choices.   
 Walk over each basic block looking for potentially hoistable
 expressions, nothing gets hoisted from the entry block.   
     Examine each expression that is very busy at the exit of this
     block.  These are the potentially hoistable expressions.   
             Current expression.   
             Number of occurrences of EXPR that can be hoisted to BB.   
             Occurrences reachable from BB.   
             We want to insert the expression into BB only once, so
             note when we've inserted it.   
             If an expression is computed in BB and is available at end of
             BB, hoist all occurrences dominated by BB to BB.   
                     An occurrence might've been already deleted
                     while processing a dominator of BB.   
             We've found a potentially hoistable expression, now
             we look at every block BB dominates to see if it
             computes the expression.   
                 Ignore self dominance.   
                 We've found a dominated block, now see if it computes
                 the busy expression and whether or not moving that
                 expression to the "beginning" of that block is safe.   
                 An occurrence might've been already deleted
                 while processing a dominator of BB.   
                   Adjust MAX_DISTANCE to account for the fact that
                   OCCR won't have to travel all of DOMINATED, but
                   only part of it.   
                 Note if the expression should be hoisted from the dominated
                 block to BB if it can reach DOMINATED unimpared.

                 Keep track of how many times this expression is hoistable
                 from a dominated block into BB.   
             If we found more than one hoistable occurrence of this
             expression, then note it in the vector of expressions to
             hoist.  It makes no sense to hoist things which are computed
             in only one BB, and doing so tends to pessimize register
             allocation.  One could increase this value to try harder
             to avoid any possible code expansion due to register
             allocation issues; however experiments have shown that
             the vast majority of hoistable expressions are only movable
             from two successors, so raising this threshold is likely
             to nullify any benefit we get from code hoisting.   
                 If (hoistable != vec::length), then there is
                 an occurrence of EXPR in BB itself.  Don't waste
                 time looking for LCA in this case.   
                       Punt, it's better to hoist these occurrences to
                       LCA.   
               Punt, no point hoisting a single occurrence.   
                 Increase register pressure of basic blocks to which
                 expr is hoisted because of extended live range of
                 output.   
                 Restore register pressure and live_in info for basic
                 blocks recorded in hoisted_bbs when expr will not be
                 hoisted.   
             Walk through occurrences of I'th expressions we want
             to hoist to BB and make the transformations.   
                 Create a pseudo-reg to store the result of reaching
                 expressions into.  Get the mode for the new pseudo
                 from the mode of the original destination pseudo.

                 It is important to use new pseudos whenever we
                 emit a set.  This will allow reload to use
                 rematerialization for such registers.   
static void insert_expr_in_table ( rtx  x,
enum machine_mode  mode,
rtx  insn,
int  antic_p,
int  avail_p,
int  max_distance,
struct hash_table_d table 
)
static

Insert expression X in INSN in the hash TABLE. If it is already present, record it as the last occurrence in INSN's basic block.

MODE is the mode of the value X is being stored into. It is only used if X is a CONST_INT.

ANTIC_P is nonzero if X is an anticipatable expression. AVAIL_P is nonzero if X is an available expression.

MAX_DISTANCE is the maximum distance in instructions this expression can be moved.

 Do not insert expression in table if it contains volatile operands,
 or if hash_expr determines the expression is something we don't want
 to or can't handle.   
     If the expression isn't found, save a pointer to the end of
     the list.   
       This is the first pattern that hashed to this index.   
       Add EXPR to end of this hash chain.   
     Set the fields of the expr element.   
 Now record the occurrence(s).   
       Found another instance of the expression in the same basic block.
       Prefer the currently recorded one.  We want the first one in the
       block and the block is scanned from start to end.   
         First occurrence of this expression in this basic block.   
         Found another instance of the expression in the same basic block.
         Prefer this occurrence to the currently recorded one.  We want
         the last one in the block and the block is scanned from start
         to end.   
         First occurrence of this expression in this basic block.   

Referenced by record_opr_changes().

static void insert_insn_end_basic_block ( )
static

Add EXPR to the end of basic block BB.

This is used by both the PRE and code hoisting.

 If the last insn is a jump, insert EXPR in front [taking care to
 handle cc0, etc. properly].  Similarly we need to care trapping
 instructions in presence of non-call exceptions.   
     FIXME: What if something in cc0/jump uses value set in new insn?   
 Likewise if the last insn is a call, as will happen in the presence
 of exception handling.   
     Keeping in mind targets with small register classes and parameters
     in registers, we search backward and place the instructions before
     the first parameter is loaded.  Do this for everyone for consistency
     and a presumption that we'll get better code elsewhere as well.   
     Since different machines initialize their parameter registers
     in different orders, assume nothing.  Collect the set of all
     parameter registers.   
     If we found all the parameter loads, then we want to insert
     before the first parameter load.

     If we did not find all the parameter loads, then we might have
     stopped on the head of the block, which could be a CODE_LABEL.
     If we inserted before the CODE_LABEL, then we would be putting
     the insn in the wrong basic block.  In that case, put the insn
     after the CODE_LABEL.  Also, respect NOTE_INSN_BASIC_BLOCK.   

References expr::bitmap_index, BLOCK_FOR_INSN, dump_file, emit_insn_after(), emit_insn_before(), expr::expr, expr_equiv_p(), gcc_assert, gcc_unreachable, gen_move_insn(), GET_CODE, INSN_UID, NULL_RTX, PATTERN, expr::reaching_reg, REG_P, REGNO, SET, SET_DEST, SET_SRC, validate_change(), XVECEXP, and XVECLEN.

static void invalidate_any_buried_refs ( )
static

Make sure there isn't a buried reference in this pattern anywhere. If there is, invalidate the entry for it since we're not capable of fixing it up just yet.. We have to be sure we know about ALL loads since the aliasing code will allow all entries in the ld_motion list to not-alias itself. If we miss a load, we will get the wrong value since gcse might common it and we won't know to fix it up.

Invalidate it in the list.

 Recursively process the insn.   

References execute(), execute_rtl_pre(), and gate_rtl_pre().

Referenced by one_code_hoisting_pass().

static bool is_too_expensive ( )
static

Return true if the graph is too expensive to optimize. PASS is the optimization about to be performed.

Trying to perform global optimizations on flow graphs which have a high connectivity will take a long time and is unlikely to be particularly useful.

In normal circumstances a cfg should have about twice as many edges as blocks. But we do not want to punish small functions which have a couple switch statements. Rather than simply threshold the number of blocks, uses something with a more graceful degradation.

 If allocating memory for the dataflow bitmaps would take up too much
 storage it's better just to disable the optimization.   

Referenced by pre_insert_copies().

static struct ls_expr* ldst_entry ( )
staticread
Here we provide the things required to do store motion towards the exit.
In order for this to be effective, gcse also needed to be taught how to
move a load when it is killed only by a store to itself.
    int i;
    float a[10];

    void foo(float scale)
    {
      for (i=0; i<10; i++)
        a[i] *= scale;
    }
'i' is both loaded and stored to in the loop. Normally, gcse cannot move
the load out since its live around the loop, and stored at the bottom
of the loop.

  The 'Load Motion' referred to and implemented in this file is
an enhancement to gcse which when using edge based LCM, recognizes
this situation and allows gcse to move the load out of the loop.

  Once gcse has hoisted the load, store motion can then push this
load towards the exit, and we end up with no loads or stores of 'i'
in the loop.      

This will search the ldst list for a matching expression. If it doesn't find one, we create one and initialize it.

Referenced by one_code_hoisting_pass().

static int load_killed_in_block_p ( const_basic_block  bb,
int  uid_limit,
const_rtx  x,
int  avail_p 
)
static

Return nonzero if the expression in X (a memory reference) is killed in block BB before or after the insn with the LUID in UID_LIMIT. AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills before UID_LIMIT.

To check the entire block, set UID_LIMIT to max_uid + 1 and AVAIL_P to 0.

 If this is a readonly then we aren't going to be changing it.   
     Ignore entries in the list that do not apply.   
     If SETTER is a call everything is clobbered.  Note that calls
     to pure functions are never put on the list, so we need not
     worry about them.   
     SETTER must be an INSN of some kind that sets memory.  Call
     note_stores to examine each hunk of memory that is modified.   

References expr::expr, expr_equiv_p(), GOBNEW, expr::hash, hash_expr(), expr::next_same_hash, NULL, hash_table_d::size, and hash_table_d::table.

rtl_opt_pass* make_pass_rtl_hoist ( )
rtl_opt_pass* make_pass_rtl_pre ( )
static void mems_conflict_for_gcse_p ( rtx  dest,
const_rtx  setter,
void *  data 
)
static

DEST is the output of an instruction. If it is a memory reference and possibly conflicts with the load found in DATA, then communicate this information back through DATA.

If DEST is not a MEM, then it will not conflict with the load. Note that function calls are assumed to clobber memory, but are handled elsewhere.

 If we are setting a MEM in our list of specially recognized MEMs,
 don't mark as killed this time.   

References oprs_unchanged_p().

static int one_code_hoisting_pass ( )
static

Top level routine to perform one code hoisting (aka unification) pass

Return nonzero if a change was made.

 Return if there's nothing to do, or it is too expensive.   
 Calculate register pressure for each basic block.   
 We need alias.   
 We are finished with alias.   

References alloc_INSN_LIST(), can_assign_to_reg_without_clobbers_p(), find_reg_equal_equiv_note(), FOR_BB_INSNS, GET_CODE, ls_expr::invalid, invalidate_any_buried_refs(), ldst_entry(), ls_expr::loads, MEM_P, NONDEBUG_INSN_P, NULL_RTX, PATTERN, REG_NOTE_KIND, REG_P, SET, SET_DEST, SET_SRC, simple_mem(), ls_expr::stores, and XEXP.

static int one_pre_gcse_pass ( )
static

Top level routine to perform one PRE GCSE pass.

Return nonzero if a change was made.

 Return if there's nothing to do, or it is too expensive.   
 We need alias.   
 We are finished with alias.   

References BB_DATA, bitmap_bit_p, bitmap_clear_bit(), BLOCK_FOR_INSN, edge_def::dest, DF_REF_INSN, DF_REF_INSN_INFO, DF_REF_NEXT_REG, DF_REF_REAL_REG, DF_REG_USE_CHAIN, EXIT_BLOCK_PTR, FOR_EACH_EDGE, get_regno_pressure_class(), NONDEBUG_INSN_P, NULL, REGNO, and basic_block_def::succs.

static int oprs_anticipatable_p ( )
static

Return nonzero if the operands of expression X are unchanged from the start of INSN's basic block up to but not including INSN.

static int oprs_available_p ( )
static

Return nonzero if the operands of expression X are unchanged from INSN to the end of INSN's basic block.

static int oprs_unchanged_p ( )
static

Return nonzero if the operands of expression X are unchanged from the start of INSN's basic block up to but not including INSN (if AVAIL_P == 0), or from INSN to the end of INSN's basic block (if AVAIL_P != 0).

If we are about to do the last recursive call needed at this level, change it into iteration. This function is called enough to be worth it.

References mem_conflict_info::conflict, and mem_conflict_info::mem.

Referenced by mems_conflict_for_gcse_p(), and oprs_unchanged_p().

static int pre_delete ( )
static

Delete redundant computations. Deletion is done by changing the insn to copy the `reaching_reg' of the expression into the result of the SET. It is left to later passes (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.

Return nonzero if a change is made.

       We only need to search antic_occr since we require ANTLOC != 0.   
           We only delete insns that have a single_set.   
               Create a pseudo-reg to store the result of reaching
               expressions into.  Get the mode for the new pseudo from
               the mode of the original destination pseudo.   

References sbitmap_vector_alloc().

static int pre_edge_insert ( )
static

Insert partially redundant expressions on edges in the CFG to make the expressions fully redundant.

 Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
 if it reaches any of the deleted expressions.   
               Now look at each deleted occurrence of this expression.   
                   Insert this expression on this edge if it would
                   reach the deleted occurrence in BB.   
                       We can't insert anything on an abnormal and
                       critical edge, so we insert the insn at the end of
                       the previous block. There are several alternatives
                       detailed in Morgans book P277 (sec 10.5) for
                       handling this situation.  This one is easiest for
                       now.   

References expr::avail_occr, BLOCK_FOR_INSN, occr::copied_p, occr::deleted_p, occr::insn, INSN_DELETED_P, occr::next, NULL, pre_expr_reaches_here_p(), and pre_insert_copy_insn().

static int pre_expr_reaches_here_p ( )
static

The wrapper for pre_expr_reaches_here_work that ensures that any memory allocated for that function is returned.

Referenced by pre_edge_insert().

static int pre_expr_reaches_here_p_work ( basic_block  occr_bb,
struct expr expr,
basic_block  bb,
char *  visited 
)
static

PRE utilities Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach block BB.

VISITED is a pointer to a working buffer for tracking which BB's have been visited. It is NULL for the top-level call.

We treat reaching expressions that go through blocks containing the same reaching expression as "not reaching". E.g. if EXPR is generated in blocks 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block 2 as not reaching. The intent is to improve the probability of finding only one reaching expression and to reduce register lifetimes by picking the closest such expression.

Has predecessor has already been visited?   
     Does this predecessor generate this expression?   


         Is this the occurrence we're looking for?
         Note that there's only one generating occurrence per block
         so we just need to check the block number.   
     Ignore this predecessor if it kills the expression.   
     Neither gen nor kill.   
 All paths have been checked.   
static int pre_gcse ( )
static

Perform GCSE optimizations using PRE. This is called by one_pre_gcse_pass after all the dataflow analysis has been done.

This is based on the original Morel-Renvoise paper Fred Chow's thesis, and lazy code motion from Knoop, Ruthing and Steffen as described in Advanced Compiler Design and Implementation.

??? A new pseudo reg is created to hold the reaching expression. The nice thing about the classical approach is that it would try to use an existing reg. If the register can't be adequately optimized [i.e. we introduce reload problems], one could add a pass here to propagate the new register through the block.

??? We don't handle single sets in PARALLELs because we're [currently] not able to copy the rest of the parallel when we insert copies to create full redundancies from partial redundancies. However, there's no reason why we can't handle PARALLELs in the cases where there are no partial redundancies.

 Compute a mapping from expression number (`bitmap_index') to
 hash table entry.   
 Delete the redundant insns first so that
 - we know what register to use for the new insns and for the other
   ones with reaching expressions
 - we know which insns are redundant when we go to create copies   
 In other places with reaching expressions, copy the expression to the
 specially allocated pseudo-reg that reaches the redundant expr.   

Referenced by pre_insert_copies().

static void pre_insert_copies ( )
static

Copy available expressions that reach the redundant expression to `reaching_reg'.

 For each available expression in the table, copy the result to
 `reaching_reg' if the expression reaches a deleted one.

 ??? The current algorithm is rather brute force.
 Need to do some profiling.   
       If the basic block isn't reachable, PPOUT will be TRUE.  However,
       we don't want to insert a copy here because the expression may not
       really be redundant.  So only insert an insn if the expression was
       deleted.  This test also avoids further processing if the
       expression wasn't deleted anywhere.   
       Set when we add a copy for that expression.   
               No need to handle this one if handled already.   
               Don't handle this one if it's a redundant one.   
               Or if the expression doesn't reach the deleted one.   
               Copy the result of avail to reaching_reg.   

References _, add_noreturn_fake_exit_edges(), alloc_gcse_mem(), alloc_hash_table(), alloc_pre_mem(), changed, compute_hash_table(), compute_ld_motion_mems(), compute_pre_data(), current_function_name(), dump_file, dump_hash_table(), edge_list, end_alias_analysis(), expr_hash_table, free_edge_list(), free_gcse_mem(), free_hash_table(), free_ld_motion_mems(), free_pre_mem(), gcc_obstack_init, gcse_obstack, init_alias_analysis(), is_too_expensive(), last_basic_block, n_basic_blocks, hash_table_d::n_elems, NULL, NUM_FIXED_BLOCKS, pre_gcse(), remove_fake_exit_edges(), and trim_ld_motion_mems().

static void pre_insert_copy_insn ( )
static

Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG. Given "old_reg <- expr" (INSN), instead of adding after it reaching_reg <- old_reg it's better to do the following: reaching_reg <- expr old_reg <- reaching_reg because this way copy propagation can discover additional PRE opportunities. But if this fails, we try the old way. When "expr" is a store, i.e. given "MEM <- old_reg", instead of adding after it reaching_reg <- old_reg it's better to add it before as follows: reaching_reg <- old_reg MEM <- reaching_reg.

 This block matches the logic in hash_scan_insn.   
     Search through the parallel looking for the set whose
     source was the expression that we're interested in.   
             If the source was a REG_EQUAL or REG_EQUIV note, we
             may not find an equivalent expression, but in this
             case the PARALLEL will have a single set.   
     Check if we can modify the set destination in the original insn.   
     Check if we can modify the set source in the original insn.   

Referenced by pre_edge_insert().

static void print_ldst_list ( )
static

Dump debugging info about the ldst list.

References MAX_GCSE_MEMORY, n_basic_blocks, n_edges, SBITMAP_ELT_TYPE, SBITMAP_SET_SIZE, and warning().

static rtx process_insert_insn ( )
static

Generate RTL to copy an EXPR to its `reaching_reg' and return it.

 Copy the expression to make sure we don't have any sharing issues.   
 If the expression is something that's an operand, like a constant,
 just copy it to a register.   
 Otherwise, make a new insn to compute this expression and make sure the
 insn will be recognized (this also adds any needed CLOBBERs).   

References expr::bitmap_index, dump_file, basic_block_def::index, and INDEX_EDGE_SUCC_BB.

static void prune_expressions ( )
static

Remove certain expressions from anticipatable and transparent sets of basic blocks that have incoming abnormal edge. For PRE remove potentially trapping expressions to avoid placing them on abnormal edges. For hoisting remove memory references that can be clobbered by calls.

         Note potentially trapping expressions.   
           Note memory references that can be clobbered by a call.
           We do not split abnormal edges in hoisting, so would
           a memory reference get hoisted along an abnormal edge,
           it would be placed /before/ the call.  Therefore, only
           constant memory references can be hoisted along abnormal
           edges.   
               Constant memory reference, e.g., a PIC address.   
             ??? Optimally, we would use interprocedural alias
             analysis to determine if this mem is actually killed
             by this call.   
     If the current block is the destination of an abnormal edge, we
     kill all trapping (for PRE) and memory (for hoist) expressions
     because we won't be able to properly place the instruction on
     the edge.  So make them neither anticipatable nor transparent.
     This is fairly conservative.

     ??? For hoisting it may be necessary to check for set-and-jump
     instructions here, not just for abnormal edges.  The general problem
     is that when an expression cannot not be placed right at the end of
     a basic block we should account for any side-effects of a subsequent
     jump instructions that could clobber the expression.  It would
     be best to implement this check along the lines of
     should_hoist_expr_to_dom where the target block is already known
     and, hence, there's no need to conservatively prune expressions on
     "intermediate" set-and-jump instructions.   

References bitmap_ior(), bitmap_not(), and basic_block_def::index.

static void prune_insertions_deletions ( )
static

It may be necessary to insert a large number of insns on edges to make the existing occurrences of expressions fully redundant. This routine examines the set of insertions and deletions and if the ratio of insertions to deletions is too high for a particular expression, then the expression is removed from the insertion/deletion sets.

N_ELEMS is the number of elements in the hash table.

 We always use I to iterate over blocks/edges and J to iterate over
 expressions.   
 Counts for the number of times an expression needs to be inserted and
 number of times an expression can be removed as a result.   
 Set of expressions which require too many insertions relative to
 the number of deletions achieved.  We will prune these out of the
 insertion/deletion sets.   
 Iterate over the edges counting the number of times each expression
 needs to be inserted.   
 Similarly for deletions, but those occur in blocks rather than on
 edges.   
 Now that we have accurate counts, iterate over the elements in the
 hash table and see if any need too many insertions relative to the
 number of evaluations that can be removed.  If so, mark them in
 PRUNE_EXPRS.   
 Now prune PRE_INSERT_MAP and PRE_DELETE_MAP based on PRUNE_EXPRS.   
static void record_last_mem_set_info ( )
static

Record memory modification information for INSN. We do not actually care about the memory location(s) that are set, or even how they are set (consider a CALL_INSN). We merely need to record which insns modify memory.

load_killed_in_block_p will handle the case of calls clobbering everything.

static void record_last_reg_set_info ( )
static

Record register first/last/block set information for REGNO in INSN.

first_set records the first place in the block where the register is set and is used to compute "anticipatability".

last_set records the last place in the block where the register is set and is used to compute "availability".

last_bb records the block for which first_set and last_set are valid, as a quick test to invalidate them.

Referenced by record_last_set_info().

static void record_last_set_info ( )
static

Called from compute_hash_table via note_stores to handle one SET or CLOBBER in an insn. DATA is really the instruction in which the SET is taking place.

Ignore pushes, they clobber nothing.

References clear_modify_mem_tables().

static int should_hoist_expr_to_dom ( basic_block  expr_bb,
struct expr expr,
basic_block  bb,
sbitmap  visited,
int  distance,
int *  bb_size,
enum reg_class  pressure_class,
int *  nregs,
bitmap  hoisted_bbs,
rtx  from 
)
static

Determine if the expression EXPR should be hoisted to EXPR_BB up in flow graph, if it can reach BB unimpared. Stop the search if the expression would need to be moved more than DISTANCE instructions.

DISTANCE is the number of instructions through which EXPR can be hoisted up in flow graph.

BB_SIZE points to an array which contains the number of instructions for each basic block.

PRESSURE_CLASS and NREGS are register class and number of hard registers for storing EXPR.

HOISTED_BBS points to a bitmap indicating basic blocks through which EXPR is hoisted.

FROM is the instruction from which EXPR is hoisted.

It's unclear exactly what Muchnick meant by "unimpared". It seems to me that the expression must either be computed or transparent in every block in the path(s) from EXPR_BB to BB. Any other definition would allow the expression to be hoisted out of loops, even if the expression wasn't a loop invariant.

Contrast this to reachability for PRE where an expression is considered reachable if any path reaches instead of all paths.

Record old information of basic block BB when it is visited
at the first time.   
 Terminate the search if distance, for which EXPR is allowed to move,
 is exhausted.   


         Prefer to hoist EXPR if register pressure is decreased.   
         Let EXPR be hoisted through basic block at no cost if one
         of following conditions is satisfied:

         1. The basic block has low register pressure.
         2. Register pressure won't be increases after hoisting EXPR.

         Constant expressions is handled conservatively, because
         hoisting constant expression aggressively results in worse
         code.  This decision is made by the observation of CSiBE
         on ARM target, while it has no obvious effect on other
         targets like x86, x86_64, mips and powerpc.   
     Not killed.   
     If EXPR can be hoisted to expr_bb, record basic blocks through
     which EXPR is hoisted in hoisted_bbs.   
         Record the basic block from which EXPR is hoisted.   

References expr::antic_occr, occr::deleted_p, find_occr_in_bb(), gcc_assert, occr::insn, and NONDEBUG_INSN_P.

Referenced by free_code_hoist_mem().

static int simple_mem ( )
static

Load Motion for loads which only kill themselves. Return true if x, a MEM, is a simple access with no side effects. These are the types of loads we consider for the ld_motion list, otherwise we let the usual aliasing take care of it.

If we are handling exceptions, we must be careful with memory references that may trap. If we are not, the behavior is undefined, so we may just continue.

 Do not consider function arguments passed on stack.   

Referenced by one_code_hoisting_pass().

static void trim_ld_motion_mems ( )
static

Remove any references that have been either invalidated or are not in the expression list for pre gcse.

     Delete if entry has been made invalid.   
         Delete if we cannot find this mem in the expression list.   
         Set the expression field if we are keeping it.   
 Show the world what we've found.   

Referenced by pre_insert_copies().

static int update_bb_reg_pressure ( )
static

Update register pressure for BB when hoisting an expression from instruction FROM, if live ranges of inputs are shrunk. Also maintain live_in information if live range of register referred in FROM is shrunk.

Return 0 if register pressure doesn't change, otherwise return the number by which register pressure is decreased.

NOTE: Register pressure won't be increased in this function.

The live range of register is shrunk only if it isn't:

  1. referred on any path from the end of this block to EXIT, or
  2. referred by insns other than FROM in this block.
     Decrease register pressure and update live_in information for
     this block.   

Referenced by add_label_notes().

static void update_ld_motion_stores ( )
static

This routine will take an expression which we are replacing with a reaching register, and update any stores that are needed if that expression is in the ld_motion list. Stores are updated by copying their SRC to the reaching register, and then storing the reaching register into the store location. These keeps the correct value in the reaching register for the loads.

     We can try to find just the REACHED stores, but is shouldn't
     matter to set the reaching reg everywhere...  some might be
     dead and should be eliminated later.   
     We replace (set mem expr) with (set reg expr) (set mem reg)
     where reg is the reaching reg used in the load.  We checked in
     compute_ld_motion_mems that we can replace (set mem expr) with
     (set reg expr) in that insn.   
         If we've already copied it, continue.   
         un-recognize this pattern since it's probably different now.   
static int want_to_gcse_p ( )
static

See whether X, the source of a set, is something we want to consider for GCSE.

 GCSE'ing constants:

 We do not specifically distinguish between constant and non-constant
 expressions in PRE and Hoist.  We use set_src_cost below to limit
 the maximum distance simple expressions can travel.

 Nevertheless, constants are much easier to GCSE, and, hence,
 it is easy to overdo the optimizations.  Usually, excessive PRE and
 Hoisting of constant leads to increased register pressure.

 RA can deal with this by rematerialing some of the constants.
 Therefore, it is important that the back-end generates sets of constants
 in a way that allows reload rematerialize them under high register
 pressure, i.e., a pseudo register with REG_EQUAL to constant
 is set only once.  Failing to do so will result in IRA/reload
 spilling such constants under high register pressure instead of
 rematerializing them.   
       Do not PRE constants.   
     FALLTHRU  
       PRE doesn't implement max_distance restriction.   

Variable Documentation

sbitmap* ae_kill
static

For available exprs

sbitmap* antloc
static

Nonzero for expressions that are locally anticipatable in the block.

bitmap blocks_with_calls
static

Bitmap indexed by block numbers to record which blocks contain function calls.

int bytes_used
static

Various variables for statistics gathering. Memory used in a pass. This isn't intended to be absolutely precise. Its intent is only to keep an eye on memory usage.

vec<modify_pair>* canon_modify_mem_list
static

This array parallels modify_mem_list, except that it stores MEMs being set and their canonicalized memory addresses.

sbitmap* comp
static

Nonzero for expressions that are computed (available) in the block.

Referenced by determine_offset(), determine_use_iv_cost_address(), dfs_next_edge(), suitable_component_p(), TB_get_command(), and zero_ssa_operands().

basic_block curr_bb
static
int curr_reg_pressure[N_REG_CLASSES]
static

Current register pressure for each pressure class.

basic_block current_bb
static
struct target_gcse default_target_gcse

Partial redundancy elimination / Hoisting for RTL. Copyright (C) 1997-2013 Free Software Foundation, Inc.

This file is part of GCC.

GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version.

GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see http://www.gnu.org/licenses/. TODO

  • reordering of memory allocation and freeing to be more space efficient
  • calc rough register pressure information and use the info to drive all kinds of code motion (including code hoisting) in a unified way. References searched while implementing this.

Compilers Principles, Techniques and Tools Aho, Sethi, Ullman Addison-Wesley, 1988

Global Optimization by Suppression of Partial Redundancies E. Morel, C. Renvoise communications of the acm, Vol. 22, Num. 2, Feb. 1979

A Portable Machine-Independent Global Optimizer - Design and Measurements Frederick Chow Stanford Ph.D. thesis, Dec. 1983

A Fast Algorithm for Code Movement Optimization D.M. Dhamdhere SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988

A Solution to a Problem with Morel and Renvoise's Global Optimization by Suppression of Partial Redundancies K-H Drechsler, M.P. Stadel ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988

Practical Adaptation of the Global Optimization Algorithm of Morel and Renvoise D.M. Dhamdhere ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991

Efficiently Computing Static Single Assignment Form and the Control Dependence Graph R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991

Lazy Code Motion J. Knoop, O. Ruthing, B. Steffen ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI

What's In a Region? Or Computing Control Dependence Regions in Near-Linear Time for Reducible Flow Control Thomas Ball ACM Letters on Programming Languages and Systems, Vol. 2, Num. 1-4, Mar-Dec 1993

An Efficient Representation for Sparse Sets Preston Briggs, Linda Torczon ACM Letters on Programming Languages and Systems, Vol. 2, Num. 1-4, Mar-Dec 1993

A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion K-H Drechsler, M.P. Stadel ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993

Partial Dead Code Elimination J. Knoop, O. Ruthing, B. Steffen ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994

Effective Partial Redundancy Elimination P. Briggs, K.D. Cooper ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994

The Program Structure Tree: Computing Control Regions in Linear Time R. Johnson, D. Pearson, K. Pingali ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994

Optimal Code Motion: Theory and Practice J. Knoop, O. Ruthing, B. Steffen ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994

The power of assignment motion J. Knoop, O. Ruthing, B. Steffen ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI

Global code motion / global value numbering C. Click ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI

Value Driven Redundancy Elimination L.T. Simpson Rice University Ph.D. thesis, Apr. 1996

Value Numbering L.T. Simpson Massively Scalar Compiler Project, Rice University, Sep. 1996

High Performance Compilers for Parallel Computing Michael Wolfe Addison-Wesley, 1996

Advanced Compiler Design and Implementation Steven Muchnick Morgan Kaufmann, 1997

Building an Optimizing Compiler Robert Morgan Digital Press, 1998

People wishing to speed up the code here should read: Elimination Algorithms for Data Flow Analysis B.G. Ryder, M.C. Paull ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986

How to Analyze Large Programs Efficiently and Informatively D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI

People wishing to do something different can find various possibilities in the above papers and elsewhere. We support GCSE via Partial Redundancy Elimination. PRE optimizations are a superset of those done by classic GCSE.

Two passes of copy/constant propagation are done around PRE or hoisting because the first one enables more GCSE and the second one helps to clean up the copies that PRE and HOIST create. This is needed more for PRE than for HOIST because code hoisting will try to use an existing register containing the common subexpression rather than create a new one. This is harder to do for PRE because of the code motion (which HOIST doesn't do).

Expressions we are interested in GCSE-ing are of the form (set (pseudo-reg) (expression)). Function want_to_gcse_p says what these are.

In addition, expressions in REG_EQUAL notes are candidates for GCSE-ing. This allows PRE to hoist expressions that are expressed in multiple insns, such as complex address calculations (e.g. for PIC code, or loads with a high part and a low part).

PRE handles moving invariant expressions out of loops (by treating them as partially redundant).

We used to support multiple passes but there are diminishing returns in doing so. The first pass usually makes 90% of the changes that are doable. A second pass can make a few more changes made possible by the first pass. Experiments show any further passes don't make enough changes to justify the expense.

A study of spec92 using an unlimited number of passes: [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83, [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2, [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1

It was found doing copy propagation between each pass enables further substitutions.

This study was done before expressions in REG_EQUAL notes were added as candidate expressions for optimization, and before the GIMPLE optimizers were added. Probably, multiple passes is even less efficient now than at the time when the study was conducted.

PRE is quite expensive in complicated functions because the DFA can take a while to converge. Hence we only perform one pass.

The steps for PRE are:

1) Build the hash table of expressions we wish to GCSE (expr_hash_table).

2) Perform the data flow analysis for PRE.

3) Delete the redundant instructions

4) Insert the required copies [if any] that make the partially redundant instructions fully redundant.

5) For other reaching expressions, insert an instruction to copy the value to a newly created pseudo that will reach the redundant instruction.

The deletion is done first so that when we do insertions we know which pseudo reg to use.

Various papers have argued that PRE DFA is expensive (O(n^2)) and others argue it is not. The number of iterations for the algorithm to converge is typically 2-4 so I don't view it as that expensive (relatively speaking).

PRE GCSE depends heavily on the second CPROP pass to clean up the copies we create. To make an expression reach the place where it's redundant, the result of the expression is copied to a new register, and the redundant expression is deleted by replacing it with this new register. Classic GCSE doesn't have this problem as much as it computes the reaching defs of each register in each block and thus can try to use an existing register. GCSE global vars.

bool doing_code_hoisting_p = false
static

Doing code hoisting.

struct hash_table_d expr_hash_table
static

Expression hash table.

Referenced by pre_insert_copies().

int flag_rerun_cse_after_global_opts

Set to non-zero if CSE should run after all GCSE optimizations are done.

int gcse_create_count
static

Number of copy instructions created.

struct obstack gcse_obstack
static

An obstack for our working variables.

Referenced by pre_insert_copies().

int gcse_subst_count
static

GCSE substitutions made.

sbitmap* hoist_vbein
static

Code Hoisting variables and subroutines. Very busy expressions.

sbitmap* hoist_vbeout
static
vec<rtx>* modify_mem_list
static

Array, indexed by basic block number for a list of insns which modify memory within that block.

bitmap modify_mem_list_set
static
sbitmap* pre_delete_map
static

Nonzero for expressions which should be deleted in a specific block.

sbitmap* pre_insert_map
static

Nonzero for expressions which should be inserted on a specific edge.

struct ls_expr* pre_ldst_mems = NULL
static

Head of the list of load/store memory refs.

hash_table<pre_ldst_expr_hasher> pre_ldst_table
static

Hashtable for the load/store memory refs.

sbitmap* pre_optimal
static

Nonzero for expressions where this block is an optimal computation point.

sbitmap* pre_redundant
static

Nonzero for expressions which are redundant in a particular block.

regset reg_set_bitmap
static

Bitmap containing one bit for each register in the program. Used when performing GCSE to track which registers have been set since the start of the basic block.

rtx test_insn
static

Used internally by can_assign_to_reg_without_clobbers_p.

sbitmap* transp
static

Compute PRE+LCM working variables. Local properties of expressions. Nonzero for expressions that are transparent in the block.